I spent 20 hours installing every plugin in the DeepSeek Harness registry — 93.9% installed cleanly, only 48.6% actually ran

[zhousm666]

I run a small site called dsh.so that independently indexes and verifies DeepSeek Harness plugins. After dsh-v0.1.3-alpha.1 was released, I did something kind of stubborn. I spent about 20 hours going through the plugins I index, installing every single one in a sandbox, then trying to actually run each one. Across 3 dsh versions, on the same L1→L5 ladder, from metadata checks to sandbox install, web boot, HTTP, plugin inventory. 13,325 plugins in the registry, verification batch started 2026-09-06. Here's the data.

Signal	Count	Share of registry
Installs cleanly (L4)	12,568	94.3%
Actually runs as a plugin (L5)	6,448	48.4%
Never reached run verification	6,755	—

I stared at that for a second when I first saw it. 13,325 plugins in the registry, 12,568 installed cleanly, only 6,448 actually ran as plugins. Of the tested registry, 6,755 never even reached run verification.

Installing is not the same as running. That sentence is basically why a verification layer matters. On dsh.so, only an L5 pass counts as verified: sandbox install, web boot, HTTP responds, plugin shows up active in the inventory. All four steps.

A few honest caveats, because details matter more than headline numbers.

About 3,009 records failed verification — mostly version windows, missing install commands, environment mismatches. Not "bad plugins".

About 3,746 install fine but never declare dsh.bundle. dsh installs them as plain dependencies and never mounts them as plugins. The report counts them separately, as neither pass nor fail.

Verified counts dip after each new dsh release, then climb back as the retest queue catches up. That's version-scoped bookkeeping, not plugins breaking.

The full report is a online page, re-aggregated from sandbox records at every build.
EN: https://www.dsh.so/reports/install-vol1/
中文: https://www.dsh.so/zh/reports/install-vol1/
methodology: https://www.dsh.so/verification/
and per-plugin install records are all open at https://www.dsh.so/data/install/

[zhousm666]

我做了个独立收录 DeepSeek Harness 插件的站点，叫 dsh.so。上个发布周期结束以后，我干了一件挺轴的事，花了大概 20 个小时，把自己收录的插件在沙箱里挨个装了一遍，再挨个试着跑起来。跨了 3 个 dsh 版本，走的是同一条 L1 到 L5 的验证阶梯，从元数据检查一路到沙箱安装、web 启动、HTTP 请求、插件清单确认。注册表一共 13,325 个插件，验证批次是 2026-09-06 启动的。数据是这样的

指标	数量	占注册表比例
能干净装上（L4）	12,568	94.3%
真能当插件跑起来（L5）	6,448	48.4%
从未到达运行验证	6,755	—

我第一次看到这组数的时候，自己先愣了一会儿。注册表里 13,325 个插件，12,568 个装得上，最后只有 6,448 个真正跑起来，已测注册表里还有 6,755 个从来没走到运行验证这一步。

装得上和跑得起来，是两回事。 这句话大概就是为什么要做验证层的原因。在 dsh.so 上，只有 L5 通过才算 verified，沙箱安装、web 启动、HTTP 通、插件清单里确实激活了它，四步全过才算数。

几句实话也放在这儿，细节比数字重要。

大概 3,009 条记录没通过验证，绝大多数是版本窗口对不上、缺安装命令、环境不匹配，不是插件本身有问题。

大概 3,746 个能装上，但自始至终没有声明 dsh.bundle，dsh 会把它们当普通依赖装上，不会挂载成插件。报告里把它们单列，既不算过，也不算挂。

每次 dsh 发新版本，verified 的数量都会先掉一截，等复测队列追上来再爬回去。这是按版本记账的正常波动，不是插件集体坏了。

完整报告是一个在线页面，每次构建都会从沙箱记录发布。
英文版在 https://www.dsh.so/reports/install-vol1/
中文版在 https://www.dsh.so/zh/reports/install-vol1/
验证方法写在 https://www.dsh.so/verification/
逐条插件的安装记录也全部开放，https://www.dsh.so/data/install/ 可以随便翻。